Cindy S. Y. Lim

Microseismic monitoring is often used at subsurface industrial sites, such as hydraulic fracturing for shale gas, stimulation of engineered geothermal systems, mature hydrocarbon reservoirs, and carbon capture and storage. Microseismic observations can be used to map growing fractures in the subsurface, to understand the geomechanical response of subsurface reservoirs to industrial operations, and to identify and characterize nascent induced seismicity. Downhole microseismic datasets are acquired using multistation geophone strings placed in deep boreholes at reservoir depths, with sampling rates at 1 kHz or higher. Artificial intelligence (AI) tools have driven a step-change in the processing of surface-based seismological data. However, similar improvements are yet to be realized for downhole microseismic processing tasks. To date, a major barrier in the development of AI tools for downhole microseismic monitoring has been the lack of publicly available data with which to train models. We seek to eliminate this barrier by curating an AI-ready downhole microseismic benchmark database (AMBER) consisting of waveforms and labels (arrival time picks) drawn from 10 downhole microseismic datasets. AMBER contains over 4,000 events with over 150,000 individual P and S wave picks. All picks have been manually verified. The database also contains over 10,000 examples of signal-free noise, which have also been manually verified. The database is provided in an AI-ready format, with Python functions for data ingestion and data augmentation. It is hoped that AMBER will stimulate and support the development of AI tools for downhole microseismic processing.

The 1st April and 9th May 2026 ML ≥ 4.0 earthquakes in central Alberta were felt across the city of Edmonton. The events were located near to Carbon Capture and Storage (CCS) operations at the Quest site. In this study we use an established induced seismicity assessment framework, applied with respect to the Quest CCS site as well as other oilfield activities in the area, to assess whether the events were induced. We found that these earthquakes were caused by CO2 injection at Quest, making this the first known example of induced seismicity caused by CCS operations. We have inverted regional seismic monitoring data to compute event depths and source mechanisms, finding that the events were located in the upper portion of crystalline basement, with oblique strike-slip/reverse focal mechanisms. We apply several induced seismicity magnitude forecasting schemes to the earthquake sequence in order to evaluate whether the seismic hazard posed by these events could have been identified. We found that two models – the seismogenic index model and a model based on extreme value theory – would both have forecasted the ML 4.0 events prior to their occurrence. These results demonstrate that induced event forecasting methods can be used to support operational decision-making. We conclude by considering options for ongoing induced seismicity hazard mitigation, both at the Quest CCS site, and more generally at all future CCS operations.

DL-based earthquake detection can detect substantially more induced microseismic events than conventional detection methods, revealing finer-scale patterns of hydraulic fracturing-induced seismicity (HFIS). We implement a novel workflow for the Preston New Road-1z (PNR-1z) shale gas site, UK, that integrates DL phase picking, Oct-Tree-based phase association, and probabilistic non-linear location to generate a high-resolution “deep” seismic catalogue from a single high-frequency (2000 Hz) borehole array. PhaseNet has detected 11,021 previously undetected events (49,472 events total), compared to the previous catalogue, in which 38,452 events were identified by a beamforming method. We find that the enhanced catalogue increases event cluster densities by an average of 42% and lowers the magnitude of completeness (M _c ) from -0.2 to -0.5, adding thousands of additional events to Gutenberg-Richter b-value analyses. The spatial patterns suggest that HFIS interacted with a fault or fracture zone north of the PNR wells, contrasting with previous event locations and their interpretations. Combining the largest and best recorded events, using locations from our independently derived catalogue, with 3D seismic reflection data reveals clustering along previously unmapped seismic discontinuities ~300 m north of the injection wells, indicating strong structural controls on HFIS. In addition, we re-estimate moment magnitudes (M _w ) using continuous downhole waveforms, finding that magnitudes reported in the previous operational catalogue were systematically underestimated by an average of 0.93 units. The temporal b-values revealed higher b-values during active injection, consistent with tensile HF-dominated microseismicity, and relatively lower b-values during pauses and certain mid-stages, indicating increased shear failure and interaction with a pre-existing fracture zone. The deep catalogue allows for more robust b-value estimation and temporal tracking, potentially reflecting evolving source characteristics and fluid-driven processes during stimulation. This study demonstrates that DL-enhanced catalogues can refine subsurface interpretations, identify re-activated structures efficiently, and contribute to improved hazard assessments in geo-energy systems.

Deep learning (DL) phase picking models have proven effective in processing large volumes of seismic data, including successfully detecting earthquakes missed by other standard detection methods. Despite their success, the applicability of existing extensively-trained DL models to high-frequency borehole datasets is currently unclear. In this study, we compare four established models (GPD, U-GPD, PhaseNet and EQTransformer) trained on regional earthquakes recorded at surface stations (100 Hz) in terms of their picking performance on high-frequency borehole data (2000 Hz) from the Preston New Road (PNR) unconventional shale gas site, in the United Kingdom (UK). The PNR-1z dataset, which we use as a benchmark, consists of continuously recorded waveforms containing over 38,000 seismic events previously catalogued, ranging in magnitudes from -2.8 to 1.1. Remarkably, all four DL models can detect induced seismicity in high frequency borehole data and two might satisfy the monitoring requirements of some users without any modifications. In particular, PhaseNet and U-GPD demonstrate exceptional recall rates of 95% and 76.6%, respectively, and detect a substantial number of new events (over 15,800 and 8,300 events, respectively). PhaseNet’s success might be attributed to its exposure to more extensive and diverse instrument dataset during training, as well as its relatively small model size, which might mitigate overfitting to its training set. U-GPD outperforms PhaseNet during periods of high seismic rates due to its smaller window size (400-samples compared to PhaseNet’s 3000-sample window). These models start missing events below M_w -0.5, suggesting that the models could benefit from additional training with microseismic datasets. Nonetheless, PhaseNet may satisfy some users’ monitoring requirements without further modification, detecting over 52,000 events at PNR. This suggests that DL models can provide efficient solutions to the big data challenge of downhole monitoring of hydraulic-fracturing induced seismicity as well as improved risk mitigation strategies at unconventional exploration sites.

Deep learning (DL) phase picking models have proven effective in processing large volumes of seismic data, including successfully detecting earthquakes missed by other standard detection methods. Despite their success, the applicability of existing extensively-trained DL models to high-frequency borehole datasets is currently unclear. In this study, we compare four established models (GPD, U-GPD, PhaseNet and EQTransformer) trained on regional earthquakes recorded at surface stations (100 Hz) in terms of their picking performance on high-frequency borehole data (2000 Hz) from the Preston New Road (PNR) unconventional shale gas site, in the United Kingdom (UK). The PNR-1z dataset, which we use as a benchmark, consists of continuously recorded waveforms containing over 38,000 seismic events previously catalogued, ranging in magnitudes from -2.8 to 1.1. Remarkably, three of the four DL models recall a good fraction of the events and two might satisfy the monitoring requirements of some users without any modifications. In particular, PhaseNet and U-GPD demonstrate exceptional recall rates of 95% and 76.6%, respectively, and detect a substantial number of new events (over 15,800 and 8,300 events, respectively). PhaseNet’s success might be attributed to its exposure to more extensive and diverse instrument dataset during training, as well as its relatively small model size, which might mitigate overfitting to its training set. U-GPD outperforms PhaseNet during periods of high seismic rates due to its smaller window size (400-samples compared to PhaseNet’s 3000-sample window). All models start missing events below M _w -0.5, suggesting that the models could benefit from additional training with microseismic datasets. Nonetheless, PhaseNet may satisfy some users' monitoring requirements without further modification, detecting over 52,000 events at PNR. This suggests that DL models can provide efficient solutions to the big data challenge of downhole monitoring of hydraulic-fracturing induced seismicity as well as improved risk mitigation strategies at unconventional exploration sites.

A document by Cindy Lim. Click on the document to view its contents.